4. 主题报告

4.1 无人驾驶的环境感知与理解 (jian yang, NJUST)

1. outline

   • 无人驾驶发展简介
     • 遥控驾驶，自主驾驶，南理工无人车，
   • 行车环境视觉感知与理解 (具体介绍贴图片)
     • 阴影检测与去除
     • 车道线检测
     • 行人检测与姿态估计
     • 场景分割与深度估计

2. 具体，如图

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

3. https://blog.csdn.net/qq_15698613/article/details/89303060

4.2：Learning to track and segment objects in videos

• 很迷的一个报告，没啥干货

4.3 AI破晓——机遇与挑战 (陶大程)

• 没听

4.4 深度学习处理器 (陈云ji)

• 没意思，贴图待定

4.5 基于知识驱动的行为理解

1. outline

   • knowledge engine - a possible direction: HAKE
   • pose - open the door of activity understanding: Alphapose, Crowdpose
   • sequence modeling: Deep RNN: semi-couple prociple
   • summary

2. why activity understanding is difficult ?

   • huge semantic Noise (compare to object recognition)
   • Long-tail distribution, few-shot problem (DL fails)
   • 结论：pose is not enough, we need konwledge pose

3. Human activity konwledge engine (HAKE)

   • to see/parse/understand the activity
   • knowledge engine construction: 见图片
   • reasoning via part states(HAKE): 见图片
   • human-object interaction
     • 见图片，几个 HOI Dataset 有：AVA, ActivityNet, Kinetics
   • conclusion:
     • activity data is semantically noisy
     • knowledge at body part can help to denote
     • HAKE:
     • HAKE based Two-stage paradigm，见图片

4. pose - open the door of activity understanding: Alphapose, Crowdpose

   • 没记录

5. sequence modeling: Deep RNN: semi-couple prociple

   • 没记录

6. summary

   • 没记录，等他主页公布PPT吧

7. 部分图片，

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

   

4.6 人工智能与未来出行

• 没学术性，贴图待定

4.7 计算机视觉的下一步：迈向大AI (罗杰波)

• 没注意，贴图待定

4.8 梯度之谜 (孟德宇)

1. issue
   • limitations of model-driven methodology
     • generally with nonconvex model
     • only fit one unsupervised image
     • slow prediction speed
   • limitations of data-driven methodology
     • require supervised-data
     • black box issue: interpretability
     • network parameters/structure are hard/easy to be designed
2. 从梯度角度思考，解决上述问题
3. 贴图片，有一些论文